刘辉, 陈文豪,等.基于二次分解协同TCNN-GRU的短期风电功率预测[J].河南理工大学学报（自然科学版）,2026,45(5):11-19.

LIU H, CHEN W H ,et al. Short-term wind power prediction based on quadratic decomposition cooperative TCNN-GRU[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(5):11-19.

基于二次分解协同TCNN-GRU的短期风电功率预测

刘辉, 陈文豪

湖北工业大学 湖北省电网智能控制与装备工程技术研究中心，湖北 武汉 430068

摘要:目的 为了提高风电功率预测精度，保证电网调度及其安稳运行，提出一种基于二次分解协同TCNN-GRU的短期风电功率预测模型。  方法 首先，使用自适应噪声的完全集合经验模态分解（complete ensemble empirical mode decomposition with adaptive noise， CEEMDAN）和变分模态分解（variational mode decomposition， VMD）构建二次分解模块对原始风电信号进行分解，降低原始风电信号的复杂性；其次，对分解后的所有分量用双路卷积神经网络（two-way convolutional neural network， TCNN）进行特征提取，挖掘各分量间的特征关系；最后，引入门控循环单元（gate recurrent unit， GRU）对所有风电功率子分量进行时序建模，累加重组后得到预测结果。  结果 以风电场采集间隔为15 min的实测数据进行实验，结果表明，与GA-BP，PSO-ELM和CNN-GRU相比，TCNN-GRU模型在不同样本量的数据集上表现均为最优。对本文模型做消融实验可知，与单次分解相比，经二次分解后预测结果的RMSE和MAE分别下降了16.49%和5.22%， R²提升了0.02；与没有结合双路卷积神经网络的CEEMDAN-VMD-GRU模型相比，预测结果的RMSE和MAE >分别下降了44.59%和62.17%，R²提升了0.03。为验证本文模型在不同季节情况下的适用性与准确性，将风电场全年数据划分为四季做对比实验，误差计算结果表明，本文模型在不同季节的表现均最优。  结论 本文模型设计合理，采用的二次分解方法和TCNN-GRU能有效提升预测精度，且在不同气候条件下表现良好，具有较强的适用性。

关键词:二次分解;双路卷积神经网络;门控循环单元;短期风电功率预测

doi:10.16186/j.cnki.1673-9787.2025100033

基金项目:国家自然科学基金资助项目（61903129）

收稿日期:2025/10/17

修回日期:2025/12/08

出版日期:2026-07-24

Short-term wind power prediction based on quadratic decomposition cooperative TCNN-GRU

Liu Hui, Chen Wenhao

Hubei Provincial Grid Intelligent Control and Equipment Engineering Technology Research Center，Hubei University of Technology，Wuhan 430068，Hubei， China

Abstract: Objectives In order to improve the accuracy of wind power prediction and ensure the dispatch and stable operation of the power grid， a short-term wind power prediction model based on quadratic decomposition cooperative TCNN-GRU was proposed.  Methods First， a quadratic decomposition module， constructed using complete ensemble empirical mode decomposition with adaptive noise （CEEMDAN） and variational mode decomposition （VMD）， was used to decompose the original wind signal and reduce its complexity. Then， all the decomposed components were extracted for features using a two-way convolutional neural network （TCNN） to explore the feature relationships between the components. Finally， the gated recurrent unit （GRU） was introduced to model the time series of all wind power sub-components， and the prediction results were obtained through cumulative reorganization.  Results Experiments were performed on measured data from wind farms with a sampling interval of 15 min. Compared with GA-BP， PSO-ELM and CNN-GRU， the TCNN-GRU model proposed in this paper performed optimally on datasets with different sample sizes. Ablation experiments on the model in this paper showed that， compared with the single decomposition， the RMSE and MAE values of the prediction results after the secondary decomposition decreased by 16.49% and 5.22%， respectively， and the R² improved by 0.02. Compared with the CEEMDAN-VMD-GRU that did not incorporate the two-way convolutional neural network， the RMSE and MAE values of its prediction results decreased by 44.59% and 62.17%， respectively， and the R² improved by 0.03. In order to verify the applicability and accuracy of the model of this paper in different seasons， the annual data of wind farms were divided into four seasons for comparison experiments， and the error calculation results showed that its performance in different seasons was optimal.  Conclusions Comprehensive results of the above experiments showed that the method in this paper was reasonable in design， and the quadratic decomposition method and TCNN-GRU could effectively improve the prediction accuracy and perform well under different climatic conditions， which was highly practical.

Key words:quadratic decomposition;convolutional neural network;gated recurrent unit;short-term wind power prediction